Home / Curated / Algorithms / FedRecon (Federated Reconstruction for Partially Local Federated Learning)

FedRecon (Federated Reconstruction for Partially Local Federated Learning)

Algorithms AIMech:FedReconFederatedReconstructionForPartiallyLocalFederatedLearning DRAFT

Aliases: FedRecon, Federated Reconstruction, Stateless Partially Local Federated Learning via Reconstruction

A stateless partially local federated learning framework (Singhal, Sidahmed, Garrett, Wu, Rush, & Prakash 2021, Sections 3-4, Equations 1-2, Algorithm 1, Proposition 4.1) that partitions model parameters into shared global parameters g in G and private local parameters l_i in L without requiring clients to persist local state across rounds: in each round t, each selected client i in S^{(t)} splits its local dataset D_i into a reconstruction support set R_i and an update query set U_i, initializes local parameters l_i^{(t,0)} fresh, performs k_r gradient descent steps on R_i with frozen g^{(t)} to reconstruct l_i^{(t)} = R(D_i, g^{(t)}, L_i) = argmin_{l_i in L} L_i(g^{(t)}, l_i) (Equation 1), performs k_u gradient descent steps on U_i with frozen reconstructed l_i^{(t)} to update g_i^{(t)}, discards l_i^{(t)}, and returns only the global parameter update Delta_i^{(t)} = g^{(t)} - g_i^{(t)} to the server (Algorithm 1).

Identity Notes & Concept Scope Distinct from stateful partially local methods such as FedPer (AIMech:FedPerFederatedLearningWithPersonalizationLayers, Arivazhagan et al. 2019), FedRep (AIMech:FedRepExploitingSharedRepresentationsForPersonalizedFederatedLearning, Collins et al. 2021), and LG-FedAvg (AIMech:LgFedAvgLocalGlobalRepresentationsFederatedLearning, Liang et al. 2020), all of which require clients to persistently store their private parameters across communication rounds and cannot directly evaluate on unseen stateless cross-device clients without a reconstruction step; FedRecon discards local parameters after every round and reconstructs them on the fly via support/query splitting, naturally generalizing to unseen clients at inference time via ReconEval.

Claims & Grounded Evidence

4 Claims

Mechanism #r481-stateless-on-the-fly-reconstruction-and-reconevaluation-mechanism

Instead of maintaining persistent client state l_i across rounds, FedRecon reconstructs l_i^{(t)} from scratch on each active client's reconstruction split R_i via k_r gradient steps with frozen global parameters g^{(t)} before updating g_i^{(t)} on update split U_i and discarding l_i^{(t)}; at inference time (ReconEval), any unseen client reconstructs its personal parameters l_i from R_i in k_r steps and evaluates on U_i (Singhal et al. 2021, Section 3, Algorithm 1, pp. 3-4).

supports Source: Federated Reconstruction: Partially Local Federated Learning · Locator: Section 3 (Equations 1-2, Algorithm 1), pp. 3-4

Scope #r481-first-order-maml-equivalence-and-bilevel-convergence-scope

Proposition 4.1 proves that for one reconstruction step (k_r = 1 on R_i with step size alpha_r) and one global update step (k_u = 1 on U_i), FedRecon optimizes the bi-level meta-learning objective min_g (1/|C|) sum_i F_i(g) = L_i(g, l_i - alpha_r nabla_l L_i(g, l_i)) using the first-order MAML gradient nabla_g L_{U_i}(g, l_i^{(t)}) (omitting the expensive second-order Hessian cross-term alpha_r nabla_g nabla_l L_i) and converges to an O(alpha_r)-stationary point at rate O(1/sqrt(T) + alpha_r) in smooth non-convex settings (Singhal et al. 2021, Section 4 & Appendix C).

supports Source: Federated Reconstruction: Partially Local Federated Learning · Locator: Section 4 (Proposition 4.1), pp. 5-6, and Appendix C (Proposition C.4), pp. 14-16

Application #r481-federated-matrix-factorization-and-next-word-prediction-application

Applied to federated collaborative filtering via matrix factorization on MovieLens-1M (where user embeddings U_i are reconstructed locally and item embeddings V are aggregated globally) and personalized next-word prediction on StackOverflow with user-specific vocabulary bias/head layers (Singhal et al. 2021, Section 5, Figures 2-4, Tables 1-2).

supports Source: Federated Reconstruction: Partially Local Federated Learning · Locator: Section 5.1-5.2 (Figures 2-4, Tables 1-2), pp. 6-9

Limitation #r481-support-split-requirement-for-reconstruction-limitation

Requires every participating and evaluated client to hold a labeled support subset R_i with enough local examples to reconstruct l_i^{(t)} in k_r steps without overfitting before global updates or inference can be performed (Singhal et al. 2021, Section 3 & Section 5.1, pp. 4 & 7).

supports Source: Federated Reconstruction: Partially Local Federated Learning · Locator: Section 3 (Dataset Splitting), p. 4, and Section 5.1, p. 7

⚠️ Curated Limitations & Negative Results

Requires splitting each client's local dataset D_i into separate reconstruction (R_i) and update (U_i) subsets, reducing the sample size available for each phase.
Performs k_r reconstruction gradient steps in every round from scratch, adding local compute overhead compared to stateful partially local methods.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:FederatedLearning` Federated Learning	broad_match	FedRecon is a stateless partially local federated learning algorithm within the broader FederatedLearning category.

Evidence Sources & Primary Literature

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`r481-primary-pdf`	Federated Reconstruction: Partially Local Federated Learning Karan Singhal, Hakim Sidahmed, Zachary Garrett, Shanshan Wu, Keith Rush, Sushant Prakash	primary paper full_text	ID: https://arxiv.org/abs/2102.03448 sha256:ebc1ae9b3dba…
`r481-arxiv-abs`	Federated Reconstruction: Partially Local Federated Learning (arXiv abstract) Karan Singhal, Hakim Sidahmed, Zachary Garrett, Shanshan Wu, Keith Rush, Sushant Prakash	primary paper abstract_only	ID: https://arxiv.org/abs/2102.03448 sha256:2450f6b72c01…
`r481-metadata-cr`	Federated Reconstruction: Partially Local Federated Learning (Crossref metadata) Karan Singhal, Hakim Sidahmed, Zachary Garrett, Shanshan Wu, Keith Rush, Sushant Prakash	primary paper abstract_only	ID: https://arxiv.org/abs/2102.03448 sha256:fc08ed8b13db…

Ontology & Grounding Topology

Open in interactive network graph →

Curation Provenance

Curator: Antigravity

Method: ai_assisted

Model: Gemini

Created: 2026-10-19T09:24:00Z

Updated: 2026-10-19T09:24:00Z

Source File: data/curated/FedReconFederatedReconstructionForPartiallyLocalFederatedLearning.yaml · Schema: curated_concept.yaml

Explore across X-Mech

Related computational & biological Mechs: Embedding Map Network Graph ProteinTraitsMech (ESM-2 pLM) DUFMech (Domain pLM) TraitMech CultureMech All 14 X-Mech Repositories